A custom multi-touch attribution system is a way to inspect the customer activity your business can observe before an order. It becomes useful only when that activity can be tied back to a dependable order record and a decision the team will actually make.
That is a different goal from recreating every influence on a buyer. Browser controls, device changes, unclicked media, and incomplete event capture mean the record will always have limits. A credible build makes those limits inspectable instead of burying them inside a scoring rule.
Begin with the order record
Choose the system that will serve as the order record of reference, then map the events and identifiers that can be compared with it. For an ecommerce business, that typically means identifying the purchase, checkout, session, customer, and campaign fields that can be joined without assuming they are more complete than they are.
The question is not whether every record will match perfectly. The question is whether the differences can be explained. If an attribution view has a materially different last-click total from the comparable OMS view, investigate the handoff before you ask the model to allocate earlier credit.
Four requirements worth checking first
1. An event record that survives the journey
Identify where visits, product views, checkout, purchase, and campaign parameters are captured or lost. A model cannot repair an event that disappears during a payment-provider handoff.
2. Useful, imperfect identity continuity
Returning activity does not need to be perfect to be informative. It does need enough first-party continuity that a return visit is not automatically treated as a completely new person. Document where that continuity breaks.
3. Stable campaign and channel definitions
Campaign taxonomy is part of the model input. If naming conventions changed, a channel was reclassified, or UTM coverage is inconsistent, record it before drawing a conclusion from the trend.
4. A named investigation
“Which touchpoints appear before orders from new customers?” is a usable investigation. “What is the exact truth of attribution?” is not. The first question determines which slices, time windows, and comparisons deserve implementation effort.
Build the model around the comparison that can challenge it, not the credit rule that makes the cleanest chart.
Use the attribution rule as a documented convention
Once the record is stable enough to investigate, last-touch, first-touch, linear, time-decay, and custom rules can each have a role. Read what attribution-model rules can actually tell you before letting any one rule become a budget mandate.
If the business needs to know whether a spend change created more total orders, an MTA path remains observational. Use it to frame the hypothesis, then rely on an incrementality test when one can create a credible comparison. For broader portfolio allocation questions, an MMM may supply the relevant lens.
What a responsible handoff includes
- The documented event and order sources, including ownership and known gaps.
- The identifier and campaign-taxonomy assumptions behind every important join.
- Reconciliation checks against the order record and other relevant reporting views.
- The attribution rules, their intended operating questions, and their boundaries.
- A list of questions that should become a test, a foundation repair, or a broader model instead.
That is the difference between a custom system and a custom dashboard. The system preserves the path from result back to evidence. See the custom MTA service page for the implementation scope.